materials-chemistry/ai produced the result/Journal of Chemical Information and Modeling 2024 · v2
Neural network predicts atomic charges from bond lists, without atomic coordinates
Chemists trained small neural networks to predict how electric charge is shared between the atoms in a molecule, using only the pattern of bonds. The networks stood in for quantum-chemistry calculations that normally need atomic positions.
spectrum · one line per step, placed by what the step does · bright lines used AI
Coordinate-Free and Low-Order Scaling Machine Learning Model for Atomic Partial Charge Prediction for Any Size of Molecules
Journal of Chemical Information and Modeling, 2024
doi:10.1021/acs.jcim.4c00376 · record aix-00208 v2 · checked 2026-10-09
- AI was for
- Property prediction
- Model family
- Multilayer perceptron
- Checked by
- Held-out
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Atoms in a molecule do not hold their electrons evenly. Some pull electron density towards themselves, leaving neighbours slightly positive. Chemists summarise this by giving each atom a partial charge, a small number that says how much of an electron it has gained or lost. These charges matter because much of chemistry is electrostatic: they help predict how molecules stick to one another, dissolve, or dock into a protein. The trouble is that partial charges are not directly measurable. They are worked out from quantum-mechanical calculations of where the electrons sit, and there are several competing recipes for dividing the electron cloud up between atoms, each giving somewhat different numbers.
Those calculations are also slow, and they need the three-dimensional positions of every atom, which must themselves be found first. Cost grows steeply with molecular size, so large molecules become awkward. The authors set out to predict partial charges from connectivity alone: which atom is bonded to which, and by what kind of bond, with no coordinates supplied. They fitted separate models to five established charge recipes, including Mulliken, Hirshfeld, CM5 and DDEC6 charges, and to semiempirical charges supplied with a molecule library.
Where AI came in
Each molecule was treated as a graph, with atoms as nodes carrying their Pauling electronegativity, a textbook measure of electron-pulling strength, and bonds as the links between them. A fixed, non-learned procedure passed information along those links out to three bonds away, boiling each atom's surroundings down to five numbers. A feed-forward neural network, with five hidden layers of fifty nodes each, then mapped those five numbers to a charge for a single atom at a time. It was trained from scratch to match charges computed by conventional quantum chemistry and semiempirical software.
The network is the result the paper reports. It takes the place of the quantum-chemistry calculation that would otherwise produce the charges, and of the geometry step that calculation depends on. Accuracy was checked against held-out molecules the networks had not seen, with a reported root-mean-square error of 0.018e for Hirshfeld charges and 0.045e overall for the DDEC6 model. Timing runs on alkane chains of 1200 to 12,000 carbon atoms indicate cost rising in step with the number of atoms.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors built a machine-learning model that predicts atomic partial charges from molecular connectivity alone, without atomic coordinates as input. A message-passing featurizer turns the molecular graph into five normalised descriptors per atom, and a dense neural network with five 50-node hidden layers maps those descriptors to a charge for one atom at a time. Separate networks were fitted to Mulliken, Hirshfeld, CM5, DDEC6 and AMSOL semiempirical charges, using 12,124 GDB13 molecules for training and 1393 for testing, and 898,466 ZINC20 lead-like molecules for training and 100,000 for testing in the semiempirical case. The reported root-mean-square error for Hirshfeld charge prediction is 0.018e, the DDEC6 model's overall RMSE on the GDB13-based test set is 0.045e, and timing tests on alkane chains of 1200 to 12,000 carbon atoms indicate O(n) scaling.
How AI was used
Molecules were taken from GDB13 as SMILES and from ZINC20 as mol2 files; SMILES were converted with Open Babel, and reference charges were computed with Gaussian 16 at B3LYP/6-31G(d) for Mulliken, Hirshfeld and CM5, with Chargemol for DDEC6, while AMSOL semiempirical charges came with the ZINC20 set. Each molecule was mapped to a graph whose nodes carry Pauling electronegativity and whose edges encode bond type, and a non-learned message-passing featurizer propagated features over the graph with a cutoff of three bonds, producing bond, atom and self contribution terms plus mean neighbouring and self electronegativity — five features per atom, each normalised by the training-set mean and standard deviation. These five features were the input to a feed-forward network implemented in TensorFlow 2.7.4 with a 5-node input layer, five 50-node tanh dense layers and a single output node, trained from scratch for 100 epochs with Adam and a mean-absolute-error loss, at batch size 16 for the Mulliken, Hirshfeld, CM5 and DDEC6 models and 768 for the semiempirical model. At prediction time the network was run per atom and its output denormalised into a partial charge; held-out test sets were used for per-element error indicators, and the trained semiempirical model was also run over long alkane chains to measure time scaling.
The shape of the work
Structural · the record, drawn
no AI
Assemble molecular data sets
Obtaining raw data, whether by measurement, download or retrieval.
are obtained from database GDB13 with C, N, O, and H elements in the smiles formatwhere the paper describes this · verbatim
no AI
Generate lowest-energy geometries
Cleaning, filtering, normalising or labelling data already obtained.
The format conversion from smiles to xyz (which does provide atomic positions) is done by Open Babel.where the paper describes this · verbatim
no AI
Compute reference charges by DFT
Numerical or physics simulation, including where a learned surrogate replaces it.
All of the Mulliken, Hirshfeld, and CM5 charges are calculated by Gaussian 16 using the B3LYP functional and a 6-31G(d) basis set.where the paper describes this · verbatim
no AI
Build per-atom descriptors with the message-passing featurizer
Encoding data into features, descriptors, embeddings or graphs.
Five features are used to describe the atomic local environment for atom i.where the paper describes this · verbatim
AI
Train charge-prediction neural networks
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
This neural network model is trained for 100 epochs with the Adam optimizerwhere the paper describes this · verbatim
AI
Predict atomic charges
Running a trained model over new data to predict, classify or score. The AI stood in for simulation.
This model accepts feature matrix Fi as input and returns a normalized predicted charge.where the paper describes this · verbatim
no AI
Evaluate accuracy on held-out test sets
Testing outputs against ground truth.
Table 2 shows the value of indicators for each model and elementwhere the paper describes this · verbatim
AI
Measure scaling on long alkane chains
Testing outputs against ground truth.
Figure 3 shows the results of the time efficiency test on 10 alkane chains containing from 1200 to 12,000 carbon atoms.where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's reported result is the predictive charge model itself and its accuracy and scaling; the learned model produces the charges the paper is about.
The model contains a 5-node input layer, five 50-node dense layers activated by the tanh functionwhere the paper describes this · verbatim
Bleiziffer’s models achieved RMSEs of 0.029e and 0.016e in a testing set based on ZINC and ChEMBL databases.where the paper describes this · verbatim
All training and testing data sets and Python scripts are provided in 10.5281/zenodo.10149110.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of Model MullikenWhich version of the model was used is not stated.
- Version of Model HirshfeldWhich version of the model was used is not stated.
- Version of Model CM5Which version of the model was used is not stated.
- Version of Model DDEC6Which version of the model was used is not stated.
- Version of Model SEWhich version of the model was used is not stated.
- What step 8 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00208, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error